Joan Comadran

AI Engineer · Barcelona

Hire an AI engineer in Barcelona who solves problems with software

Industrial engineer turned AI engineer and full stack developer. I solve problems with software, AI included: agents with MCP, apps, websites 10x faster, and paid acquisition with proper tracking. Barcelona-based, remote-friendly, founder-mode when you need it.

Deciding whether you need an engineer, a product, or a full-time hire? The buying guide is Hire an AI engineer: how to pick one who ships. This page is the offer.

Offer

What I take on.

  • 01

    AI agents & MCP

    Production agents that call real tools: WhatsApp, CRMs, calendars, and internal APIs, via Model Context Protocol and LLM tool use. Multi-tenant where it matters, BYOK when cost control matters more.

  • 02

    AI-powered apps end-to-end

    From brief to shipped SaaS: Next.js, TypeScript, Supabase, auth, billing hooks, and model pipelines that survive contact with real users.

  • 03

    Paid acquisition + tracking

    Google and Meta campaigns with measurement that still works post-iOS14: GA4, GTM, server-side tracking, and dashboards so spend maps to outcomes instead of vibes.

What I build

Typical projects.

  • WhatsApp agents that answer and qualify leads

    Claude replies in seconds, collects the lead's details with tool use, books appointments inside WhatsApp, and escalates to a human when needed. Multi-tenant, BYOK, Meta Cloud API direct. What booking costs, and whether to build or buy, is in WhatsApp appointment booking with AI: build or buy.

    Claient case study →
  • AI inside an existing product

    A model woven into a real workflow instead of bolted on: a Claude Script Doctor and a dramatic-arc Hype Graph inside an act-based film production pipeline.

    Dairector case study →
  • AI SaaS shipped end-to-end

    SEO drafts from real product data, on-brand imagery and editorial calendars, pushed straight into WordPress and Shopify from a single workspace.

    Adorea.ai case study →
  • On-device AI with no server

    Subject masks, object removal, background removal and upscaling running in the browser with ONNX Runtime Web. Photos never leave the user's device.

    Lightframes case study →
  • Agents and tools wired to your systems

    Agents that call CRMs, calendars and internal APIs through MCP or plain function calling, plus the automations and tracking that keep a team's workflow moving. What an agent should keep between sessions is in AI agent memory: what to store, what to forget.

    MCP vs function calling →

Built for production

What keeps an agent working after the demo.

A demo makes one model call on the happy path. A production agent survives retries, timeouts, long conversations and leaked keys. These are the patterns I ship by default, each written up from real projects.

  • Idempotent webhooks

    Incoming messages are claimed before processing, so a Meta retry never sends your lead the same reply twice.

    Read the write-up →
  • Durable long-running steps

    A step log, a queue and tools that know whether they can run twice, so an agent that dies mid-task resumes instead of restarting.

    Read the write-up →
  • Bounded model cost

    A rolling context window plus summary, prompt caching on the fixed prefix and real token and USD cost logged per account.

    Read the write-up →
  • Isolated data and keys

    Row Level Security per tenant, API keys encrypted and server-only, and hard spend caps so a leaked key has a ceiling.

    Read the write-up →

Proof

Shipped, not hypothetical.

All projects →
“Many inquiries reached us on WhatsApp outside business hours: which cava to gift, prices, shipping, and they went unanswered until the next day. Now Claient replies right away and that sale no longer slips away: we have noticed an increase in bottle orders since we turned it on.”
Núria, Online Manager at Sogas Mascaró, on Claient. The winery uses it for WhatsApp inquiries and to book visits and meetings with the CEO.

On-device AI in the browser: Lightframes, my open-source photo editor, runs its AI models in the user’s browser with ONNX Runtime Web: subject masks, MI-GAN object removal, IS-Net background removal and Real-ESRGAN upscaling. No uploads, no server, works offline as a PWA.

Writing

Writing on building with AI.

All posts →

How I work

Stack & process.

Default stack

  • TypeScript
  • Next.js
  • React
  • Node.js
  • Python
  • Supabase
  • Claude / Anthropic
  • OpenAI
  • MCP
  • WhatsApp Cloud API
  • GA4 / GTM
  • Google Ads
  • Meta Ads

Flexible when the client stack requires it (PHP, WordPress, existing backends). The constant is shipping something users can touch.

  1. Week 0–1

    Scope & stack

    Clarify the outcome, constraints, and success metrics. Pick the smallest stack that can ship, usually TypeScript / Next.js plus the model and tools you already trust.

  2. Week 1–3

    Build the spine

    Core flows first: auth, data model, the agent or LLM path, and one happy-path integration. You get something clickable early, not a wall of slides.

  3. Week 3–5

    Harden & instrument

    Edge cases, observability, cost controls, and tracking. If growth is in scope, wire campaigns and attribution before scaling spend.

  4. Ongoing

    Ship & iterate

    Launch, watch real usage, then tighten prompts, tools, and funnels. I stay close through the first production weeks so nothing drifts.

FAQ

Hiring an AI engineer: quick answers.

What does an AI engineer do for a business?
Turns a business problem into working software where a model does part of the job: an agent that answers customers, a tool that drafts or classifies work, a feature inside your product. Most of the work is not the prompt but everything around it: integrations with your systems, permissions, cost control, failure handling and tracking, so it keeps working after launch.
Who is a good fit to hire you as an AI engineer?
Founders and teams that need an AI-powered app or agent shipped end-to-end, not just a prototype. Ideal when product, engineering, and (optionally) paid acquisition need to move together.
Do you build MCP and production AI agents?
Yes. Agents with tool use, WhatsApp and API integrations, multi-tenant setups, and Model Context Protocol where it fits. Claient, my WhatsApp agent that answers and qualifies leads with Claude, is one in production. The bar is production use, not a weekend demo.
How does a project start, and how long does it take?
With a clear outcome and a short discovery week to fix scope, constraints and success criteria. Then the build runs in 3–5 week slices with something usable early, and I stay close through the first production weeks. Timelines depend on scope and on the integrations involved, so I give one after the discovery week, not before.
What stack do you use?
TypeScript, Next.js, React and Node.js by default, Python when it fits, Supabase for data and auth, Claude or OpenAI for models, MCP and the WhatsApp Cloud API for integrations, and GA4, GTM, Google Ads and Meta Ads for tracking and growth. I adapt to your stack when it requires it (PHP, WordPress, existing backends).
Are you available remotely from Barcelona?
Yes. Based in Barcelona and hybrid/remote-friendly. I work in English, Spanish and Catalan, and collaboration works over email, LinkedIn, and the usual async tools.
How do you handle data privacy and API keys?
Data stays separated per customer with Row Level Security, API keys live encrypted and are only ever called from the server, and spend caps limit the damage if a key leaks. When the data shouldn't leave the device at all, models can run in the browser, as they do in Lightframes, with no uploads.
How do I contact you to start?
Email me at joan.comadran2@gmail.com or message me on LinkedIn with what you want live in the next few weeks and the problem behind it. I reply with questions, then a proposed scope.

Get in touch

Let’s ship the agent, or the product.

Tell me what you want live in the next few weeks. AI engineering, full-stack AI apps, paid acquisition rebuilds. If it’s interesting to build, I’m in.